Making AML more effective requires new approaches and new thinking – some of which may be radical. Here is a new idea that will change how we’ve constructed AML operations for the past 20 years: False positives no longer matter. They are now irrelevant.
This sounds strange – crazy to some. After all, isn’t “reducing false positives” among the most pressing issues for AML teams? Aren’t most of the marketing messages we read about some new software or processes that will “reduce false positives”?
The problem with this thinking is that it anchors AML to outdated and inefficient approaches. While many fixate on false positives, this fixation hinders the adoption of ready-made and simple-to-deploy AI solutions. These AI solutions render false positives a non-issue. And this is where it gets interesting.
Although AI will one day be capable of reducing the number of alerts generated from screening and monitoring systems, its real impact available today to every AML team is that it automates the review process for all sanction, watchlist, adverse media, and PEP alerts. AI then applies easy-to-understand reasoning models that resolve each alert in seconds, either closing it or escalating it to a person for investigation.
The False Positives Problem
To understand this radical shift in approach, let’s examine the problem of false positives more closely. False positives are born from two primary functions of AML operations: screening for potential sanctions, watchlists, adverse media, and PEP risks, as well as monitoring transactions for potentially suspicious activity. Let’s focus on the main source of false positives here—screening alerts—and discuss why two decades of efforts to reduce them have failed and why investing more time and money trying to reduce them makes little sense.
Over 90% of sanctions, watchlist, PEP, and adverse media alerts are not “false” in the sense of malfunctions. They stem from systems operating as designed—casting a wide net to ensure that true positives are not missed. Missing a true positive is a genuine problem. The current approach prioritizes sensitivity over specificity—catching everything that may need human review. In screening, a person or business name is, and will always be, a poor unique identifier, leading to endless alerts no matter how finely tuned. There’s only so much knob turning you can do before you increase the risk of missing something legitimate.
Enter the component technologies, collectively known as “AI.” These modern computing technologies include Machine Learning, Natural Language Processing, Intelligent Document Processing, Generative AI (aka LLMs), and Agentic AI. Below is a brief description of each:
- Machine Learning (ML) is a misnomer; it’s more accurate to call it “Machine Teaching,” where systems spot patterns.
- Natural Language Processing (NLP) – Machines can read, process, and comprehend both structured (database) data and unstructured (news article) information.
- Intelligent Document Processing (IDP) – Extract, analyze, and interpret data from IDs and documents of all types.
- Generative AI (LLMs) – Classifying, understanding, and creating text and images.
- AI Agents – Applications that solve multi-step problems independently and collaboratively when necessary to achieve a specific objective or outcome.
The issue with false positives is the associated costs. Financial institutions and organizations cannot afford to hire hundreds or thousands of people to review each alert. When banks create large teams or outsource the work, many workers leave for other jobs because the tasks are so monotonous. Additionally, the repetitive nature of the work makes it prone to errors. Ironically, while many consider this AML work to be “lower level,” it can escalate into significant problems.
Worsening the problem is that “false positives” can lead to significant backlogs, which increase regulatory risk as potential matches remain undetected and deadlines for identifying and reporting suspicious activity are missed. In short, alert backlogs are arguably the most costly and risky part of AML. Yet, the approach to addressing this issue is to continue the same thing.
This is where AI is transforming AML forever.
How AI is Making False Positives Irrelevant
Imagine having the resources to hire as many high-performing alert analysts as you need, who never quit or take time off. False positives wouldn’t pose a problem. You’d be confident there would never be a risk of backlog, and every alert would be reviewed and resolved precisely as the procedures require. Of course, this isn’t a reality in a world of limited financial and human resources, but it is the reality when using AI.
The breakthrough lies in how AI replicates human data gathering and reasoning—analyzing context, cross-referencing data, and documenting decisions like the best analysts—at a scale and speed that humans cannot match.
Let’s walk through how AI works in sanction and watchlist alert screening. Typically, alerts exist because the name of the sanctioned or listed person or entity is similar to that of a customer or counterparty. This may occur when parties are involved in a payment, such as a wire transfer, or it could arise from a new customer opening an account. Alerts like this pile up every day. To determine if there is a true match – whether the person listed by OFAC is indeed the one set to receive a wire transfer passing through your bank’s system – an analyst compares the information from OFAC with the details they have on the payment’s beneficiary.
The Travel Rule requires that the beneficiary’s name and address information be included. The analyst then begins their work steps.
- Is the name an exact match, somewhat similar, or not very similar at all?
- If the name is similar, is there any other identifying information?
- Is there location information available? How specific is that information? Is it referring to both a country and a city?
- If that information exists, does it match or is it nearby?
- If it’s nearby, is there age or date of birth information available for comparison? If not, what then?
- Is there a passport or some other identifying information?
- Analysts repeat these or similar steps throughout the day at institutions worldwide. Each one takes a few minutes, and some may take over 10 minutes to determine whether the alert is a true match, a clear non-match, or requires escalation for further review.
- All this is documented in a brief, perhaps just three or four sentences, explanation that supports the analyst’s decision.
What is happening here is fully replicated by AI. The data is read and processed. When one step cannot resolve the alert, it proceeds to the next step, and so on, until the AI recommends a decision, either to close the alert or escalate it for further review. The AI documents its work, providing both an audit trail and an explanation of the reasoning it applied.
There are multiple levels of magic here. The first is that the AI does this in nanoseconds, following procedures consistently each time. There is no deviation from requirements. Additionally, if the institution wishes, the AI can easily ingest and consider more data, such as adverse media or corporate record information, ensuring that a broader range of information is used for making better decisions. Without AI, incorporating additional data is helpful, but it slows work, which must be balanced against the risk of creating a backlog. So, in addition to faster and more consistent work, AI enables more work of higher quality that without AI may be cost and risk prohibitive.
AI’s benefits extend to adverse media screening as well. Consider each step involved in reviewing new reports or other media: reading to gain understanding, identifying the name of the person or entity in question, determining if that person or entity is a focal point of the adverse report or merely an ancillary character, assessing if further investigation is needed, and documenting the outcome to incorporate the results into the case investigation. For some customers or counterparties, there may be dozens of articles.
AI, using Natural Language Processing, Machine Learning (ML), and Large Language Models (LLMs), completes each of these steps in the time it takes you to read this sentence. It can present results to the analyst, who will then determine, according to their bank’s procedures, what to do next.
AI is now accelerating the enhanced due diligence process, where gathering, organizing, and assessing information takes many hours, is prone to omission, and involves way to much copying and pasting. For transaction monitoring alerts, AI is now analyzing and offering recommended decisions on the most common alerts involving cash and wire transfers.
The Promise of AI for AML is Real
AI has changed how we think about false positives: We aren’t going to reduce them. We’re going to resolve them—all of them, in seconds.
Changes in how we approach AML have come slowly, if at all. The promise of AI is finally real for AML. This shift requires us to reconsider how we think about and discuss issues that are decades old. Reorient your mind to understand that some of our struggles are no longer a concern. This frees up time to worry about something else.
To fully appreciate how WorkFusion is transforming AML, request a demo today.
